JP7093009B2 - Optimization device, optimization device control method, and optimization device control program - Google Patents

Description

The present invention relates to an optimization device, a control method for the optimization device, and a control program for the optimization device.

Conventionally, there is a method of calculating a multivariable optimization problem, which is a weak point of a Neumann computer, by an optimization device using an Ising-type energy function (sometimes called an Ising machine or a Boltzmann machine). The optimizer replaces the problem to be calculated with the Ising model, which is a model representing the spin behavior of a magnetic material, and calculates.

The optimizer can also model the problem using, for example, a neural network. In that case, each of the multiple bits corresponding to the multiple spins included in the Ising model is a weight value (also known as a coupling coefficient) that indicates the value of the other bit and the magnitude of the interaction between the other bit and its own bit. It functions as a neuron that outputs 0 or 1 depending on (called). The optimizer finds, for example, a combination of states of each neuron in which the minimum value of the energy function value (hereinafter referred to as energy) of the Zing model is obtained by a stochastic search method such as simulated annealing as a solution.

Conventionally, there is an optimization device that calculates the combination of states of each neuron whose energy is minimized by performing simulated annealing using a digital circuit (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2018-041351 Japanese Unexamined Patent Publication No. 2016-051314 Japanese Unexamined Patent Publication No. 2006-072691

By the way, in many optimization problems, variables take three or more values instead of two values of 0 and 1, such as an integer programming problem.
However, when expressing the state of a neuron using a variable having three or more values in a conventional optimizer, whether or not to update the state of the neuron by expressing one variable using a plurality of bits is determined. There is a problem that the circuit scale for calculating the energy change of the Ising model to be determined increases.

In one aspect, the present invention provides an optimization device capable of suppressing an increase in the circuit scale of an optimization device that handles variables having three or more values, a control method for the optimization device, and a control program for the optimization device. The purpose is to do.

In one embodiment, a plurality of spins included in the singing model obtained by transforming the problem to be calculated correspond to a plurality of spins, each of which is represented by a variable of m value (m is a positive integer of 3 or more). The first neuron is based on the state change direction of the first neuron whose state has been updated and the weighting coefficient indicating the magnitude of the interaction between the first neuron and the second neuron. The state of 2 neurons is determined based on a calculation circuit that calculates the two energy changes of the Isging model due to a 2 ⁿ (n is an integer of 0 or more) changes in both positive and negative directions, a random number, and a temperature parameter. Based on the magnitude relationship between the thermal excitation energy and the two energy changes, it is determined whether or not to allow the update of the state of the second neuron that causes the two energy changes, and whether to allow the update. The second neuron, which outputs a determination result of whether or not, and has a state transition determination circuit that limits the update in which the state of the second neuron exceeds the upper limit value or falls below the lower limit value, and is different from each other. Based on the plurality of zing devices that output the determination results for each of the above-mentioned determination results and the determination results output by each of the plurality of zing devices, one of the second neurons that allow renewal is selected, and the state is described. An optimizer having an updated neuron selection circuit that selects a change direction and a state update circuit that updates the state of the second neuron selected by the update neuron selection circuit based on the state change direction is provided. To.

Also, in one embodiment, a method of controlling the optimizer is provided. Also, in one embodiment, a control program for the optimizer is provided.

In one aspect, the present invention can suppress an increase in the circuit scale of an optimizer that handles variables having three or more values.

It is a figure which shows an example of the optimization apparatus of 1st Embodiment. It is a flowchart which shows the flow of an example of the operation of the optimization apparatus of 1st Embodiment. It is a figure which shows an example of the optimization apparatus of the 2nd Embodiment. It is a figure which shows an example of the 1st ΔE change part. It is a figure which shows an example of the 2nd ΔE change part. It is a figure which shows an example of the transition possibility determination part. It is a flowchart which shows the flow of an example of the operation of the optimization apparatus of 2nd Embodiment. It is a figure which shows an example of the optimization apparatus of the 3rd Embodiment. It is a figure which shows an example of the 1st domain confirmation part. It is a figure which shows an example of the 2nd domain confirmation part. It is a flowchart which shows the flow of an example of the operation of the optimization apparatus of 3rd Embodiment. It is a figure which shows an example of the optimization apparatus of 4th Embodiment. It is a figure which shows the example of the part where the shift operation is performed in the local field update part. It is a flowchart which shows the flow of an example of the operation of the optimization apparatus of 4th Embodiment. It is a figure which shows the hardware example of a control device. It is a flowchart which shows the flow of an example of the control method of the optimization device by a control device.

Hereinafter, embodiments for carrying out the invention will be described with reference to the drawings.
The optimizer shown below is a combination of states when the energy function is the minimum value among the combinations of the states of each of the multiple neurons corresponding to multiple spins included in the Ising model that transforms the problem to be calculated. Is to search for. Each neuron is represented by a variable with an m value (m is an integer of 3 or more). Further, the variable has a predetermined bit width, and when it changes, it is assumed that the variable changes by 2 ⁿ (n is an integer of 0 or more) in the positive or negative direction.

The Ising-type energy function E (x) is defined by, for example, the following equation (1).

One item on the right side is the sum of the products of the states of the two neurons and the connection coefficient for all combinations of the two neurons that can be selected from all the neurons, without omission or duplication. x _i is a variable indicating the state of the i-th neuron (also called a state variable), x _j is a variable indicating the state of the j-th neuron, and _Wij is the interaction between the i-th and j-th neurons. It is a weighting coefficient indicating the magnitude of. In addition, _Wii = 0. In addition, W _ij = W _ji in many cases (that is, the coefficient matrix based on the weighting coefficient is often a symmetric matrix). Further, the weighting coefficient _Wij has a predetermined bit width (for example, 16 bits, 32 bits, 64 bits, 128 bits, etc.).

The two items on the right side are the sum of the products of the bias coefficients of all neurons and the variables representing the state of the neurons. b _i indicates the bias coefficient of the i-th neuron.

Further, when the variable x _i is an m value and the variable x _i does not reach the upper limit or the lower limit, the change amount Δ x _i of the variable x _i is ± 2 ⁿ .
The energy change ΔE _{i ±} accompanying the change of the variable x _i in the positive direction and the negative direction is expressed by the following equation (2).

As shown in equation (2), the energy change ΔE _{i ±} has two values depending on the state change direction (positive or negative) of the i-th neuron.
In equation (2), _hi is called a local field (local field).

(First Embodiment)
FIG. 1 is a diagram showing an example of an optimization device according to the first embodiment.
The optimization device 10 of the first embodiment includes N Ising devices 11a1, 11a2, ..., 11aN, an update neuron selection circuit 12, a state update circuit 13, and a control unit 14 that perform calculations on N neurons. .. Each of the singing devices 11a1 to 11aN may be a one-chip semiconductor integrated circuit, or the singing devices 11a1 to 11aN may be included in the one-chip semiconductor integrated circuit.

The Ising device 11a1 performs a calculation on the first neuron (neuron having identification information ID = 1) out of N neurons. The Ising device 11a1 has a calculation circuit 11b1 and a state transition determination circuit 11c1.

The calculation circuit 11b1 receives the state change direction DR of the neuron whose state has been updated and the identification information ID of the neuron from the update neuron selection circuit 12 among the N neurons. Then, the calculation circuit 11b1 has an energy change ΔE ₁₊ , ΔE ₁ based on a weighting coefficient indicating the magnitude of the interaction between the neuron identified by the identification information ID and the first neuron and the state change direction DR. _-Calculate . Energy changes ΔE ₁₊ and ΔE _1- are two energy changes in the Ising model due to changes in the state of the first neuron by 2 ⁿ in both positive and negative directions.

In the example of FIG. 1, the calculation circuit 11b1 has a register 11b11, a local field update unit 11b12, and an energy change calculation unit 11b13.
The register 11b 11 stores N weighting coefficients W ₁₁ , W ₁₂ , ..., W _1N . The weighting coefficients W ₁₁ , W ₁₂ , ..., W _1N may be stored in a memory such as a SRAM (Static Random Access Memory). It should be noted that all the weighting coefficients _W11 to _WNN may be stored in one memory.

The local field update unit 11b12 selects a weighting coefficient indicating the magnitude of the interaction between the neuron and the first neuron from the identification information ID of the neuron whose state has been updated. Then, the local field update unit _11b12 updates the local field h1 based on the selected weighting coefficient and the state change direction DR of the neuron whose state has been updated. The initial value of the local field h ₁ is, for example, a bias coefficient b ₁ .

The change in the local field h ₁ when the state of the neuron (jth neuron) changes by 2 ⁿ in the positive direction is +2 ⁿ W _1j , and the change in the local field h ₁ when the state changes by 2 ⁿ in the negative direction. Can be expressed as -2 ⁿ W _1j .

Therefore, the local field update unit 11b12 adds or subtracts the value obtained by shifting the weighting coefficient W _1j to the left by n bits to the current local field h ₁ according to the state change direction DR, thereby adding or subtracting the local field h. ₁ can be updated. Such a local field update unit 11b12 can be realized by using, for example, a selector, an adder (or subtractor), and a shift calculation circuit. When n = 0, the shift calculation circuit is unnecessary.

The energy change calculation unit 11b13 calculates the energy changes ΔE ₁₊ and ΔE _1- based on the equation (2). The energy change calculation unit 11b13 can be realized by using, for example, a shift calculation circuit and a code inversion circuit. When n = 0, the shift calculation circuit is unnecessary.

The state transition determination circuit 11c1 allows the update of the state of the first neuron that causes the energy changes ΔE ₁₊ and ΔE _1- , respectively, based on the magnitude relationship between the thermal excitation energy and the energy changes ΔE ₁₊ and ΔE _1- . Judge whether or not. Then, the state transition determination circuit 11c1 outputs a determination result as to whether or not the update is permitted. Further, the state transition determination circuit 11c1 limits the update in which the state of the first neuron exceeds the upper limit value or falls below the lower limit value. That is, the state transition determination circuit 11c1 prevents the update that is outside the definition range of the state of the first neuron. The domain of the variable representing the state of each neuron is determined by, for example, the type of problem or the hardware limitation of the optimization device 10.

The thermal excitation energy is determined based on, for example, a random number generated by a random number generation circuit described later and a temperature parameter T input from the control unit 14.
Further, the limit information LMT including the upper limit value and the lower limit value is supplied from the control unit 14. The state transition determination circuit 11c1 determines, for example, energy change ΔE ₁₊ or energy change ΔE _1- when the state of the first neuron exceeds the upper limit value or falls below the lower limit value by changing 2 ⁿ in the positive or negative direction. , Change to a given positive value that is large enough. As a result, the state transition determination circuit 11c1 prevents the update of the state of the first neuron from being allowed. This predetermined positive value may be determined based on the detection result of the value of the determination result as to whether or not to allow the update when the energy change is changed to the predetermined positive value, for example. ..

Further, the state transition determination circuit 11c1 determines whether or not to allow the update when the state of the first neuron exceeds the upper limit value or falls below the lower limit value by changing ²ⁿ in the positive or negative direction. , May be changed to a value indicating that the update is not allowed.

A circuit example of such a state transition determination circuit 11c1 will be described in the second and subsequent embodiments.
For the other Ising devices 11a2 to 11aN, the same processing as for the Ising devices 11a1 is performed for the 2nd to Nth neurons.

For example, the calculation circuit 11b2 of the Ising device 11a2 has a register 11b21, a local field update unit 11b22, and an energy change calculation unit 11b23. The register 11b21 stores N weighting coefficients W ₂₁ , W ₂₂ , ..., W _2N . The local field update unit 11b22 selects a weighting coefficient indicating the magnitude of the interaction between the neuron and the second neuron from the identification information ID of the neuron whose state has been updated. Then, the local field update unit _11b22 updates the local field h2 based on the selected weighting coefficient and the state change direction DR of the neuron whose state has been updated. The energy change calculation unit 11b23 calculates the energy changes ΔE ₂₊ and ΔE _2- based on the equation (2). Then, the state transition determination circuit 11c2 of the singing apparatus 11a2 is the state of the second neuron that brings about the energy changes ΔE ₂₊ and ΔE _2- based on the magnitude relationship between the thermal excitation energy and the energy changes ΔE ₂₊ and ΔE _2- . Determine whether to allow each update. Then, the state transition determination circuit 11c2 outputs a determination result as to whether or not the update is permitted. Further, the state transition determination circuit 11c2 limits the update in which the state of the second neuron exceeds the upper limit value or falls below the lower limit value.

The calculation circuit 11bN of the Ising device 11aN has a register 11bN1, a local field update unit 11bN2, and an energy change calculation unit 11bN3. The register 11bN1 stores N weighting coefficients W _N1 , W _N2 , ..., W _NN . The local field update unit 11bN2 selects a weighting coefficient indicating the magnitude of the interaction between the neuron whose state has been updated and the Nth neuron from the identification information ID of the update neuron. Then, the local field update unit _11bN2 updates the local field hN based on the selected weighting coefficient and the state change direction DR of the neuron whose state has been updated. The energy change calculation unit 11bN3 calculates the energy changes ΔEN ₊ and ΔE _N− based on the equation (2). Then, the state transition determination circuit 11cN of the singing apparatus 11aN is the state of the Nth neuron that brings about the energy change ΔEN ₊ , ΔE _N− based on the magnitude relationship between the thermal excitation energy and the energy change ΔEN ₊ , ΔE _N− . Determine whether to allow each update. Then, the state transition determination circuit 11cN outputs a determination result as to whether or not the update is permitted. Further, the state transition determination circuit 11cN limits the update in which the state of the Nth neuron exceeds the upper limit value or falls below the lower limit value.

The renewal neuron selection circuit 12 selects one neuron that allows renewal and selects the state change direction DR based on the determination results output by each of the Ising devices 11a1 to 11aN. The renewal neuron selection circuit 12 outputs the identification information ID of the neuron that allows the selected renewal and the state change direction DR.

The state update circuit 13 changes (updates) the state of the neuron selected by the update neuron selection circuit 12 in the positive or negative direction by 2 ⁿ based on the state change direction DR. The state update circuit 13 has a storage unit (for example, a register) and holds the states of N neurons.

The control unit 14 transmits / receives information to / from the control device 15 (for example, a PC (Personal Computer)). For example, the control unit 14 receives the weighting coefficient _{Wij expressing the Ising model from the control device 15 and sets the weighting coefficient ij for} the Ising devices 11a1 to 11aN (stored in the registers 11b11 to 11bN1). Further, the control unit 14 may receive an initial value of the local field h _i (for example, a bias coefficient bi) from the control device 15 and set it for the Ising devices 11a1 to _11aN .

Further, the control unit 14 receives annealing conditions (information on how to lower the maximum value and minimum value of the temperature parameter T, the value of the temperature parameter T, etc.) from the control device 15, and based on the annealing conditions, the Ising device 11a1 The temperature parameter T is set for ~ 11aN. The control unit 14 gradually reduces the value of the temperature parameter T based on the annealing conditions. Further, the control unit 14 receives the above-mentioned restriction information LMT from the control device 15, and sets the restriction information LMT including the upper limit value and the lower limit value of the state of each neuron for the Ising devices 11a1 to 11aN. Further, the control unit 14 solves, for example, the state of each neuron held in the state update circuit 13 when the value of the temperature parameter T is reduced a predetermined number of times (or when the temperature parameter T reaches the minimum value). Is transmitted to the control device 15.

The control unit 14 can be realized by, for example, an electronic circuit for a specific purpose such as an ASIC (Application Specific Integrated Circuit) or an FPGA. The control unit 14 may be a processor such as a CPU (Central Processing Unit) or a DSP (Digital Signal Processor). In that case, the processor executes the program stored in the memory (not shown) to perform the processing of the control unit 14 as described above. The function of the control unit 14 may be included in the control device 15.

Hereinafter, an operation example of the optimization device 10 according to the first embodiment will be shown.
FIG. 2 is a flowchart showing a flow of an example of the operation of the optimization device according to the first embodiment.

First, the control unit 14 sets the weighting coefficient _Wij received from the control device 15, the initial value of the local field hi, the initial value of the temperature parameter T, and the limit information LMT in the singing devices 11a1 to _11aN (step S1). ..

The local field update units 11b12 to _11bN2 update the local field hi by the above-mentioned process in the second and subsequent update processes (step S2).
Each of the energy change calculation units 11b13 to 11bN3 calculates two energy changes ΔE _{i +} and ΔE _i− based on the equation (2) (step S3).

Each of the state transition determination circuits 11c1 to 11cN determines whether or not to allow the update of the state of each neuron that causes the energy changes ΔE _{i +} and ΔE _i − by the above-mentioned processing (step S4). Further, in the process of step S4, each of the state transition determination circuits 11c1 to 11cN limits the update in which the state of each neuron is out of the definition range by the above-mentioned process.

The renewal neuron selection circuit 12 selects one neuron that allows renewal and selects the state change direction DR based on the determination results output by each of the Ising devices 11a1 to 11aN (step S5).

The state update circuit 13 updates the state of the neuron selected by the update neuron selection circuit 12 based on the state change direction DR (step S6).
The control unit 14 determines whether or not the update process of steps S2 to S6 has been repeated a predetermined number of times N1 (step S7). If the update process is not repeated a predetermined number of times N1, the update process of steps S2 to S6 is repeated.

When the update process is repeated a predetermined number of times N1, the control unit 14 determines whether or not the number of changes of the temperature parameter T (the number of temperature changes) has reached the predetermined number of times N2 (step S8).
When the number of temperature changes has not reached the predetermined number of times N2, the control unit 14 changes the temperature parameter T (lowers the temperature) (step S9). How to change the values of the temperature parameter T, N1 and N2 a predetermined number of times (how small the values are to be reduced at one time, etc.) is determined based on the annealing conditions. After the process of step S9, the process from step S2 is repeated.

When the number of temperature changes reaches N2 a predetermined number of times, the control unit 14 transmits (outputs) the state of each neuron held in the state update circuit 13 to the control device 15 as a solution (calculation result). (Step S10).

In the optimization device 10 of the first embodiment as described above, the permissible change in the state of each neuron having a value of 3 or more is set to ± 2 ⁿ (n is an integer of 0 or more), and the singing is performed. The configuration of the circuit that calculates the energy change of the model is simplified and the increase in the circuit scale can be suppressed. Further, by handling the state of a neuron having a value of 3 or more in a small circuit scale, the range of problems that can be solved by the optimization device 10 can be expanded.

The value of n may be different for each neuron.
(Second embodiment)
FIG. 3 is a diagram showing an example of the optimization device of the second embodiment. In FIG. 3, the same elements as those of the optimization device 10 of the first embodiment shown in FIG. 1 are designated by the same reference numerals.

In the optimization device 20 of the second embodiment, the state transition determination circuit 21b1 of the Ising device 21a1 has a ΔE change unit 21b11,21b12 and a transition possibility determination unit 21b13, 21b14.

When the variable x ₁ changes by + 2 ⁿ and exceeds the upper limit value, the ΔE changing unit 21b11 changes the energy change ΔE ₁₊ to a predetermined positive value ΔE _max .
The ΔE change unit 21b12 changes the energy change ΔE _1- to a predetermined positive value ΔE _max when the variable x ₁ changes by -2 ⁿ and falls below the lower limit.

The positive value ΔE _max is _a sufficiently large value so that the probability that the variable x1 is allowed to be updated becomes almost zero when this value is supplied to the transition possibility determination units 21b13 and 21b14.
FIG. 4 is a diagram showing an example of the first ΔE change unit. FIG. 4 shows an example of the ΔE changing unit 21b11.

The ΔE changing unit 21b11 has a ²ⁿ addition unit 30, a comparison circuit 31, and a selector 32.
_The ²ⁿ addition unit 30 adds ²ⁿ to the current variable x1 supplied from the state update circuit 13.

The comparison circuit 31 compares the upper limit value x _1max of the variable x ₁ included in the restriction information LMT with the addition result in the ²ⁿ addition unit 30. Then, the comparison circuit 31 outputs 1 when the addition result exceeds the upper limit value x _1max , and outputs 0 in other cases.

The selector 32 inputs the energy change ΔE ₁₊ and the positive value ΔE _max included in the limitation information LMT, and when the output of the comparison circuit 31 is 1, outputs the positive value ΔE _max and outputs the positive value ΔE max of the comparison circuit 31. When the output is 0, the energy change ΔE ₁₊ is output.

FIG. 5 is a diagram showing an example of the second ΔE change unit. FIG. 5 shows an example of the ΔE changing unit 21b12.
The ΔE changing unit 21b12 has a ²ⁿ subtracting unit 33, a comparison circuit 34, and a selector 35.

_The ²ⁿ subtraction unit 33 subtracts ²ⁿ from the current variable x1 supplied from the state update circuit 13.
The comparison circuit 34 compares the lower limit value x _{1 min} of the variable x ₁ included in the restriction information LMT with the subtraction result in the ²ⁿ subtraction unit 33. Then, the comparison circuit 34 outputs 1 when the subtraction result is less than the lower limit value x _{1 min} , and outputs 0 in other cases.

The selector 35 inputs the energy change ΔE _1- and a predetermined positive value ΔE _max included in the limitation information LMT, and when the output of the comparison circuit 34 is 1, outputs a positive value ΔE _max for comparison. When the output of the circuit 34 is 0, the energy change ΔE _1- is output.

The transition possibility determination unit 21b13 in FIG. 3 allows the update of the state of the first neuron that brings about the energy change ΔE ₁₊ based on the magnitude relationship between the energy change ΔE ₁₊ or the positive value ΔE _max and the thermal excitation energy. Judge whether or not.

Whether the transition possibility determination unit 21b14 allows the update of the state of the first neuron that causes the energy change ΔE ₁ based on the magnitude relationship between the energy change ΔE _1- or the positive value ΔE _max and the thermal excitation energy. Judge whether or not.

FIG. 6 is a diagram showing an example of a transition possibility determination unit.
The transition possibility determination unit 21b13 includes a code inversion unit 40, an adder 41, a random number generation circuit 42, a selection rule application unit 43, a multiplier 44, and a comparison circuit 45.

The code inversion unit 40 inverts the sign by multiplying the energy change ΔE ₁₊ or the positive value ΔE _max output by the ΔE change unit 21b11 by -1.
The adder 41 adds the offset value off generated by the offset generation circuit 23, which will be described later, to the output value of the code inversion unit 40.

The random number generation circuit 42 generates a uniform random number r of 0 or more and 1 or less.
The selection law application unit 43 outputs a value based on the selection law (metropolis method or Gibbs method) for performing simulated annealing.

When simulated annealing is performed, if the permissible probability A (ΔE, T) of the state transition that causes a certain energy change ΔE is determined by the following equation (3), the state is in the limit of infinite time (number of iterations). Has been proven to reach the optimal solution.

In equation (3), T is the temperature parameter T described above.
When the permissible probability A (ΔE, T) expressed by the equation (3) is used and the steady state is reached after sufficient repetition, the occupancy probability of each state follows the Boltzmann distribution with respect to the thermal equilibrium state in thermodynamics. Then, since the occupancy probability of the low energy state increases as the temperature is gradually lowered from the high temperature, the low energy state should be obtained when the temperature is sufficiently lowered. This method is called simulated annealing because this is very similar to the state change when the material is annealed. At this time, the probabilistic state transition in which the energy rises corresponds to thermal excitation in physics.

The circuit that outputs the flag information (= 1) indicating that the state transition that causes the energy change ΔE with the permissible probability A (ΔE, T) is allowed is uniform with f (−ΔE / T) in the equation (3). This can be realized by a comparator that outputs a value based on the comparison result with the random number r.

However, the same function can be realized by making the following transformations. Even if the same monotonic increase function is applied to two numbers, the magnitude relationship does not change. Therefore, even if the same monotonic increasing function is applied to the two inputs of the comparator, the output of the comparator does not change. For example, the inverse function f ^-1 (-ΔE / T) of f (-ΔE / T) as a monotonic increase function acting on f (-ΔE / T), and f ^-1 as a monotonic increase function acting on a uniform random number r. It is possible to use f ^-1 (r) in which −ΔE / T of (−ΔE / T) is r. In that case, it can be seen that the circuit having the same function as the above-mentioned comparator may be a circuit that outputs 1 when −ΔE / T is larger than f ^-1 (r). Further, since the temperature parameter T is positive, the circuit may be a circuit that outputs 1 when −ΔE is larger than T · f ^-1 (r).

The selection rule application unit 43 of FIG. 6 outputs the value of f ^-1 (r) by using the conversion table that converts the input uniform random number r into the value of f ^-1 (r) described above. When the metropolis method is applied, f ^-1 (r) is log (r). The conversion table is stored in a memory such as a RAM (Random Access Memory) or a flash memory.

The multiplier 44 outputs the product (T · f ^-1 (r)) of the temperature parameter T supplied from the control unit 14 and f ^-1 (r). T · f ^-1 (r) corresponds to the thermal excitation energy.
The comparison circuit 45 compares the addition result by the adder 41 with T · f ^-1 (r), and if the addition result is larger than T · f ^-1 (r), outputs 1 as flag information FLG ₁₊ . If it is smaller than T · f ^-1 (r), 0 is output as flag information FLG ₁₊ .

The transition possibility determination unit 21b14 has the same circuit configuration as the transition possibility determination unit 21b13.
Of the Ising devices 21a1 to 21aN, the elements other than the Ising device 21a1 have the same configuration as the Ising device 21a1 and perform the same processing on the 2nd to Nth neurons.

For example, the state transition determination circuit 21bN of the Ising device 21aN has a ΔE changing unit 21bN1,21bN2 and a transition possibility determination unit 21bN3,21bN4.
The ΔE change unit 21bN1 changes the energy change ΔE _{N +} to a predetermined positive value ΔE _max when the variable x _N changes by + 2 ⁿ and exceeds the upper limit value.

The ΔE change unit 21bN2 changes the energy change ΔE _N− to a predetermined positive value ΔE _max when the variable x _N changes by -2 ⁿ and falls below the lower limit.
Whether or not the transition possibility determination unit 21bN3 allows the update of the state of the Nth neuron that causes the energy change ΔEN ₊ based on the magnitude relationship between the energy change ΔEN ₊ or the positive value ΔE _max and the thermal excitation energy. Is determined.

Whether the transition possibility determination unit 21bN4 allows the update of the state of the Nth neuron that causes the energy change ΔE _N− based on the magnitude relationship between the energy change ΔE _N− or the positive value ΔE _max and the thermal excitation energy. Judge whether or not.

Similar to the update neuron selection circuit 12 shown in FIG. 1, the update neuron selection circuit 22 selects one neuron that allows update based on the determination result (flag information) output by each of the Ising devices 21a1 to 21aN. At the same time, select the state change direction DR. Then, the update neuron selection circuit 22 outputs the identification information ID of the neuron that allows the selected update and the state change direction DR. In the optimization device 20 of the second embodiment, the renewal neuron selection circuit 22 also outputs the flag information FLG for the selected renewal-allowing neuron.

The optimization device 20 of the second embodiment has an offset generation circuit 23 that generates an offset value off. The offset generation circuit 23 increases the offset value off when the flag information FLG supplied from the update neuron selection circuit 22 indicates that the state transition is not allowed (that is, when the state transition does not occur). On the other hand, the offset generation circuit 23 resets the offset value off to 0 when the flag information FLG indicates that the state transition is allowed (that is, when the state transition occurs). When the offset value off becomes large, the state transition is easily allowed, and when the current state is in the local solution, the escape from the local solution is promoted.

Hereinafter, an operation example of the optimization device 20 according to the second embodiment will be shown.
FIG. 7 is a flowchart showing a flow of an example of the operation of the optimization device according to the second embodiment.

First, the control unit 14 sets the weighting coefficient _Wij received from the control device 15, the initial value of the local field hi, the initial value of the temperature parameter T, and the limit information LMT in the rising devices _21a1 to 21aN (step S20). .. In the optimization device 20 of the second embodiment, the limit information LMT includes a positive value ΔE _max and a domain (upper limit value and lower limit value) of each variable _xi .

The processing of steps S21 and S22 is the same as the processing of steps S2 and S3 shown in FIG.
After the processing of step S22, each of the state transition determination circuits 21b1 to 21bN changes the energy change ΔE _{i +} or the energy change ΔE _i− for the variable x _i that is out of the definition range when updating to a positive value ΔE _max (. Step S23).

Then, each of the state transition determination circuits 21b1 to 21bN determines whether or not to allow the update of the state of each neuron that causes the energy changes ΔE _{i +} and ΔE _i- (step S24). The process of step S24 also includes the above-mentioned addition process of the offset value off.

Subsequent processes of steps S25 to S30 are the same as the processes of steps S5 to S10 shown in FIG.
In the optimization device 20 of the second embodiment as described above, the same effect as that of the optimization device 10 of the first embodiment can be obtained. Further, the optimization device 20 changes the energy change ΔE _{i +} or the energy change ΔE _i− for the variable x _i that is out of the definition range when updating to a positive value ΔE _max , so that the variable x _i is out of the definition range. Update is suppressed.

(Third embodiment)
FIG. 8 is a diagram showing an example of the optimization device according to the third embodiment. In FIG. 8, the same elements as those of the optimization devices 10 and 20 shown in FIGS. 1 and 3 are designated by the same reference numerals.

In the optimization device 50 of the third embodiment, the state transition determination circuit 51b1 of the Ising device 51a1 has a transition possibility determination unit 51b11, 51b12 and a domain confirmation unit 51b13, 51b14.

The transition possibility determination unit 51b11 determines whether or not to allow the update of the state of the first neuron that causes the energy change ΔE ₁₊ based on the magnitude relationship between the energy change ΔE ₁₊ and the thermal excitation energy.

The transition possibility determination unit 51b12 determines whether or not to allow the update of the state of the first neuron that causes the energy change ΔE ₁ based on the magnitude relationship between the energy change ΔE _1- and the thermal excitation energy.

The circuit configuration of the transition possibility determination units 51b11 and 51b12 is the same as the circuit configuration of the transition possibility determination unit 21b13 shown in FIG.
The domain confirmation unit 51b13 sets the flag information indicating the determination result of whether to allow the update to 0 (indicates that the update is not allowed) when the variable x ₁ changes by + 2 ⁿ and exceeds the upper limit value. ..

When the variable x ₁ changes by -2 ⁿ and falls below the lower limit, the domain confirmation unit 51b14 sets the flag information indicating the determination result of whether to allow the update to 0.
FIG. 9 is a diagram showing an example of the first domain confirmation unit. FIG. 9 shows an example of the domain confirmation unit 51b13.

The domain confirmation unit 51b13 has a ²ⁿ addition unit 60, a comparison circuit 61, and an AND (logical product) circuit 62.
_The ²ⁿ addition unit 60 adds ²ⁿ to the current variable x1 supplied from the state update circuit 13.

The comparison circuit 61 compares the upper limit value x _1max of the variable x ₁ included in the restriction information LMT with the addition result in the ²ⁿ addition unit 60. Then, the comparison circuit 61 outputs 0 when the addition result exceeds the upper limit value x _1max , and outputs 1 in other cases.

The flag information FLG _1a output by the transition possibility determination unit 51b11 is input to one input terminal of the AND circuit 62, and the output of the comparison circuit 61 is input to the other input terminal. Then, the AND circuit 62 outputs the value (0 or 1) of the flag information FLG _1a when the output of the comparison circuit 61 is 1, and the flag information FLG _1a when the output of the comparison circuit 61 is 0. 0 is output regardless of the value of.

FIG. 10 is a diagram showing an example of the second domain confirmation unit. FIG. 10 shows an example of the domain confirmation unit 51b14.
The domain confirmation unit 51b14 has a ²ⁿ subtraction unit 63, a comparison circuit 64, and an AND circuit 65.

_The ²ⁿ subtraction unit 63 subtracts ²ⁿ from the current variable x1 supplied from the state update circuit 13.
The comparison circuit 64 compares the lower limit value x _{1 min} of the variable x ₁ included in the restriction information LMT with the subtraction result in the ²ⁿ subtraction unit 63. Then, the comparison circuit 64 outputs 0 when the subtraction result is less than the lower limit value x _{1 min} , and outputs 1 in other cases.

The flag information FLG _1b output by the transition possibility determination unit 51b12 is input to one input terminal of the AND circuit 65, and the output of the comparison circuit 64 is input to the other input terminal. Then, the AND circuit 65 outputs the value (0 or 1) of the flag information FLG _1b when the output of the comparison circuit 64 is 1, and the flag information FLG _1b when the output of the comparison circuit 64 is 0. 0 is output regardless of the value of.

Of the Ising devices 51a1 to 51aN, the elements other than the Ising device 51a1 have the same configuration as the Ising device 51a1 and perform the same processing on the 2nd to Nth neurons.

For example, the state transition determination circuit 51bN of the Ising device 51aN has transition possibility determination units 51bN1, 51bN2 and a domain confirmation unit 51bN3, 51bN4.
The transition possibility determination unit 51bN1 determines whether or not to allow the update of the state of the Nth neuron that causes the energy change ΔEN ₊ based on the magnitude relationship between the energy change ΔEN ₊ and the thermal excitation energy.

The transition possibility determination unit 51bN2 determines whether or not to allow the update of the state of the Nth neuron that causes the energy change ΔE _N− based on the magnitude relationship between the energy change ΔE _N− and the thermal excitation energy.

When the upper limit value is exceeded by the variable x _N changing by +2 ⁿ , the domain confirmation unit 51bN3 sets the flag information indicating the determination result of whether or not to allow the update to 0.
When the variable x _N changes by -2 ⁿ and falls below the lower limit, the domain confirmation unit 51bN4 sets the flag information indicating the determination result of whether to allow the update to 0.

Hereinafter, an operation example of the optimization device 50 according to the third embodiment will be shown.
FIG. 11 is a flowchart showing a flow of an example of the operation of the optimization device according to the third embodiment.

First, the control unit 14 sets the weighting coefficient _Wij received from the control device 15, the initial value of the local field hi, the initial value of the temperature parameter T, and the limit information LMT in the singing devices _51a1 to 51aN (step S40). .. In the optimization device 50 of the third embodiment, the restriction information LMT includes a domain (upper limit value and lower limit value) of each variable _xi .

The processing of steps S41 and S42 is the same as the processing of steps S2 and S3 shown in FIG.
After the processing of step S42, each of the state transition determination circuits 51b1 to 51bN determines whether or not to allow the update of the state of each neuron that causes the energy changes ΔE _{i +} and ΔE _i- (step S43). The process of step S43 also includes the above-mentioned addition process of the offset value off.

After that, each of the state transition determination circuits 51b1 to 51bN sets the flag information for the variable _xi , which is out of the definition range when updating, to 0 (step S44).
Subsequent processes of steps S45 to S50 are the same as the processes of steps S5 to S10 shown in FIG.

In the optimization device 50 of the third embodiment as described above, the same effect as that of the optimization device 10 of the first embodiment can be obtained. Further, the optimization device 50 sets the flag information for the variable x _i that is out of the definition range to 0 when updating, so that the update that the variable x _i is out of the definition range is surely suppressed.

(Fourth Embodiment)
FIG. 12 is a diagram showing an example of the optimization device according to the fourth embodiment. In FIG. 12, the same elements as those of the optimization devices 10, 20, and 50 shown in FIGS. 1, 3, and 8 are designated by the same reference numerals.

The optimization device 70 of the fourth embodiment has shift amount holding units 73a1 to 73aN for holding shift amounts n1 to nN supplied to the calculation circuits 71b1 to 71bN and the state update circuit 72. The shift amounts n1 to nN are values that determine the change width of each state of the N neurons, and are values of 0 or more. For example, the change width of the state of the first neuron can be expressed as ± 2 ⁿ¹ , and the change width of the state of the Nth neuron can be expressed as ± 2 ^nN . The shift amounts n1 to nN may be different from each other.

The shift amount holding units 73a1 to 73aN are, for example, registers, SRAMs, and the like.
The shift amounts n1 to nN are supplied to each of the local field update units 71b11 to 71bN1 included in the calculation circuits 71b1 to 71bN of the Ising devices 71a1 to 71aN. Then, each of the local field update units 71b11 to 71bN1 updates the local field by using the shift amount for the neuron whose state has been updated.

FIG. 13 is a diagram showing an example of a portion where the shift operation is performed in the local field update unit. FIG. 13 shows a part of the local field update unit 71bi1 that updates the local field for the i-th neuron among the local field update units 71b11 to 71bN1.

The local field update unit 71bi1 has a mismatch circuit 80, a selection circuit 81, and a shift calculation circuit 82.
The mismatch circuit 80 outputs 1 when the identification information ID (= i) of the i-th neuron and the identification information ID (= j) of the neuron whose state has been updated do not match, and when they match. Outputs 0 to.

For example, the mismatch circuit 80 compares the identification information IDs of the i-th neuron and the j-th neuron bit by bit, outputs 0 if all the bits match, and outputs 1 otherwise. Output.

FIG. 13 shows a circuit example of the mismatch circuit 80. The mismatch circuit 80 has k ExOR (exclusive OR) circuits 80a1, 80a2, ..., 80ak and a plurality of OR (logical sum) circuits (for example, OR circuits 80b1 to 80bm, 80c).

The ExOR circuits 80a1 to 80ak compare the identification information IDs of the i-th neuron and the j-th neuron, each consisting of k bits, one bit at a time, and output 0 if they match, and 1 if they do not match. Output. For example, the ExOR circuit 80a1 compares the least significant bits (i ₁ and j ₁ ) of the identification information IDs of the i-th neuron and the j-th neuron, and outputs 0 if they match. If they are different, 1 is output. The ExOR circuit 80ak compares the most significant bits (ik and jk) of the identification information IDs of the _i -th neuron and the _j -th neuron, and outputs 0 if they match, and is different. If so, 1 is output.

The OR circuits 80b1 to 80bm receive two outputs of the ExOR circuits 80a1 to 80ak and output the logical sum of the two. For example, the OR circuit 80b1 receives the output of the ExOR circuits 80a1 and 80a2, and outputs the logical sum of the two.

Although not shown, the outputs of the OR circuits 80b1 to 80bm are input to any of the plurality of OR circuits in the next stage by two, and the outputs of the plurality of OR circuits in that stage are two by two. The process of being input to any of the plurality of OR circuits in the next stage is repeated according to the value of k. Then, the output of the OR circuit 80c at the last stage becomes the output of the mismatch circuit 80.

The selection circuit 81 selects and outputs the weighting coefficient _Wij when the output of the mismatch circuit 80 is 1, and selects and outputs 0 when the output of the mismatch circuit 80 is 0.
The shift calculation circuit 82 outputs 0 when the output of the selection circuit 81 is 0, and when the output of the selection circuit 81 is the weighting coefficient _Wij , the weighting coefficient _Wij is left-shifted by nj bits. (W _ij * (1 << nj) = 2 ^nj W _ij ) is output. The change in the local field h _i is obtained by multiplying 2 ^{nj Wij} _by 1 or -1 according to the state change direction DR.

In FIG. 12, any one of the shift amounts n1 to nN is supplied to each of the energy change calculation units 71b12 to 71bN2 included in the calculation circuits 71b1 to 71bN of the Ising devices 71a1 to 71aN. For example, the energy change calculation unit 71b12 receives the shift amount n1 and calculates ΔE 1 ₊ = -2 ⁿ¹ h ₁ and ΔE _1- = 2 ⁿ¹ h ₁ from the equation (2). The energy change calculation unit 71bN2 receives the shift amount nN and calculates ΔE _{N +} = -2 ^nN h _N and ΔE _N − = 2 ^nN h _N from the equation (2).

The state transition determination circuits 71c1 to 71cN are substantially the same as the state transition determination circuits 21b1 to 21bN of the optimization device 20 of the second embodiment or the state transition determination circuits 51b1 to 51bN of the optimization device 50 of the third embodiment. It is the same. However, in the ΔE changing unit 21b11,21b12 to 21bN1,21bN2 and the domain confirmation unit 51b13, 51b14 to 51bN3, 51bN4, addition or subtraction of ²ⁿ is performed with respect to the current neuron state as described above ( See FIGS. 4, 5, 9, and 10). Therefore, any one of the shift amounts n1 to nN is supplied as the shift amount n to each of the state transition determination circuits 71c1 to 71cN.

The shift amounts n1 to nN are supplied to the state update circuit 72, and one shift amount corresponding to the identification information ID output by the update neuron selection circuit 22 is selected. Then, the state update circuit 72 changes the state of the neuron represented by the identification information ID by 2 ^ns in the positive or negative direction based on the selected shift amount (hereinafter referred to as ns) and the state change direction DR. Let me.

The control unit 74 has the same function as the control unit 14, and sets the shift amount n1 to nN in the shift amount holding units 73a1 to 73aN. The control unit 74 may change the shift amount n1 to nN at the time of updating the state of the neuron. This makes it possible to control the convergence speed to the optimum solution. For example, when the state of all neurons does not change even after performing a predetermined number of update processes (when the current state is in the local solution), the control unit 74 increases the shift amount n1 to nN to change the energy. The absolute value increases and escape from the local solution is promoted. This makes it possible to accelerate the convergence speed.

Hereinafter, an operation example of the optimization device 70 according to the fourth embodiment will be shown.
FIG. 14 is a flowchart showing a flow of an example of the operation of the optimization device according to the fourth embodiment.

First, the control unit 74 sets the weighting coefficient _Wij received from the control device 15, the initial value of the local field hi, the initial value of the temperature parameter T, and the limit information LMT for the singing devices _71a1 to 71aN. Further, the control unit 74 sets the shift amounts n1 to nN received from the control device 15 with respect to the shift amount holding units 73a1 to 73aN (step S60). Subsequent processing in steps S61 to S69 is the same as in steps S2 to S10 shown in FIG. 2, except that the set shift amounts n1 to nN are used. The control unit 74 may change the magnitude of the shift amounts n1 to nN, for example, after the processing of step S68.

In the optimization device 70 of the fourth embodiment as described above, the same effect as that of the optimization device 10 of the first embodiment can be obtained. Further, in the optimization device 70, since the control unit 74 individually sets the shift amounts n1 to nN for each of the N neurons, the change width of the state can be changed for each neuron. Further, the control unit 74 can also change the shift amounts n1 to nN during the process of updating the state of the neuron, and can control the convergence speed to the optimum solution.

By the way, in the optimization devices 10, 20, 50, 70 of the first to fourth embodiments described above, the functions of the control units 14, 74 are such that a processor in a control device such as a PC executes a control program. It may be realized by.

FIG. 15 is a diagram showing a hardware example of the control device.
The control device 90 includes a CPU 91, a RAM 92, an HDD (Hard Disk Drive) 93, an image signal processing unit 94, an input signal processing unit 95, a medium reader 96, a communication interface 97, and an interface 98. The above unit is connected to the bus.

The CPU 91 is a processor including an arithmetic circuit that executes a program instruction. The CPU 91 loads at least a part of the programs and data stored in the HDD 93 into the RAM 92 and executes the program. The CPU 91 may include a plurality of processor cores, the control device 90 may include a plurality of processors, and the control of the optimization device 98a may be executed in parallel by using the plurality of processors or the processor cores. ..

The RAM 92 is a volatile semiconductor memory that temporarily stores a program executed by the CPU 91 and data used by the CPU 91 for calculation. The control device 90 may include a type of memory other than the RAM, or may include a plurality of memories.

The HDD 93 is a non-volatile storage device that stores software programs such as an OS (Operating System), middleware, and application software, and data. The program includes, for example, a control program that causes the control device 90 to execute a function of setting a weighting coefficient, limit information LMT, and controlling the temperature parameter T. The control device 90 may be provided with other types of storage devices such as a flash memory and an SSD (Solid State Drive), or may be provided with a plurality of non-volatile storage devices.

The image signal processing unit 94 outputs an image (for example, an image showing the calculation result of the optimization problem) to the display 94a connected to the control device 90 according to the command from the CPU 91. As the display 94a, a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD: Liquid Crystal Display), a plasma display (PDP: Plasma Display Panel), an organic EL (OEL: Organic Electro-Luminescence) display, or the like can be used. ..

The input signal processing unit 95 acquires an input signal from the input device 95a connected to the control device 90 and outputs the input signal to the CPU 91. As the input device 95a, a pointing device such as a mouse, a touch panel, a touch pad, a trackball, a keyboard, a remote controller, a button switch, or the like can be used. Further, a plurality of types of input devices may be connected to the control device 90.

The medium reader 96 is a reading device that reads programs and data recorded on the recording medium 96a. As the recording medium 96a, for example, a magnetic disk, an optical disk, a magneto-optical disk (MO: Magneto-Optical disk), a semiconductor memory, or the like can be used. The magnetic disk includes a flexible disk (FD) and an HDD. Optical discs include CDs (Compact Discs) and DVDs (Digital Versatile Discs).

The medium reader 96 copies, for example, a program or data read from the recording medium 96a to another recording medium such as the RAM 92 or the HDD 93. The read program is executed by, for example, the CPU 91. The recording medium 96a may be a portable recording medium and may be used for distribution of programs and data. Further, the recording medium 96a or HDD 93 may be referred to as a computer-readable recording medium.

The communication interface 97 is an interface that is connected to the network 97a and communicates with other information processing devices via the network 97a. The communication interface 97 may be a wired communication interface connected to a communication device such as a switch by a cable, or may be a wireless communication interface connected to a base station by a wireless link.

The interface 98 is an interface that communicates with the optimization device 98a. The optimization device 98a is, for example, the optimization devices 10, 20, 50, 70 of the first to fourth embodiments, excluding the control units 14, 74.

FIG. 16 is a flowchart showing a flow of an example of a control method of the optimization device by the control device.
In the control device 90, the control unit 74 sets the weight coefficient _Wij , the initial value of the local field _hi , the initial value of the temperature parameter T, and the limit information LMT for the optimization device 98a. Further, when the optimization device 98a includes the shift amount holding units 73a1 to 73aN as shown in FIG. 12, the control device 90 sets the shift amount n1 to nN (step S70).

After that, the control device 90 receives, for example, a signal transmitted each time the optimization device 98a performs the above-mentioned update process, and determines whether or not the update process has been repeated a predetermined number of times N1 (step S71). If the update process is not repeated a predetermined number of times N1, the process of step S71 is repeated.

When the update process is repeated a predetermined number of times N1, the control device 90 determines whether or not the number of times the temperature parameter T is changed (the number of times the temperature is changed) reaches the predetermined number of times N2 (step S72).

When the number of temperature changes has not reached the predetermined number of times N2, the control device 90 changes the temperature parameter T (lowers the temperature) (step S73). In the process of step S73, the control device 90 transmits a value of the temperature parameter T smaller than the value previously transmitted to the optimization device 98a. How to change the values of the temperature parameter T, N1 and N2 a predetermined number of times (how small the values are to be reduced at one time, etc.) is determined based on the annealing conditions. After the process of step S73, the process from step S71 is repeated. The control device 90 may change the magnitude of the shift amounts n1 to nN, for example, after the processing of step S73.

When the number of temperature changes reaches a predetermined number of times N2, the control device 90 changes the state (variable x _i (i = 1 to N)) of each neuron at that time, for example, the state update circuit 13 shown in FIG. It is acquired as the calculation result of the optimization problem from the storage unit in (step S74).

After that, the control device 90 displays the received calculation result on, for example, the display 94a shown in FIG. 15 (step S75), and ends the control of the optimization device 98a.
As described above, the processing content of the control device 90 can be realized by causing a computer to execute a program.

The program can be recorded on a computer-readable recording medium (eg, recording medium 96a). As the recording medium, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like can be used. Magnetic disks include FDs and HDDs. The optical disc includes a CD, a CD-R (Recordable) / RW (Rewritable), a DVD, and a DVD-R / RW. The program may be recorded and distributed on portable recording media. In that case, the program may be copied from the portable recording medium to another recording medium (for example, HDD93) and executed.

Although the optimization device of the present invention, the control method of the optimization device, and one viewpoint of the control program of the optimization device have been described above based on the embodiment, these are only examples and are limited to the above description. It's not something.

10 Optimization device 11a1 to 11aN Rising device 11b1 to 11bN Calculation circuit 11b11 to 11bN1 Register 11b12 to 11bN2 Local field update unit 11b13 to 11bN3 Energy change calculation unit 11c1 to 11cN State transition determination circuit 12 Update neuron selection circuit 13 State update circuit 14 Part 20 Control device

Claims (7)

Of the multiple neurons whose state is represented using variables of m value (m is a positive integer of 3 or more) corresponding to multiple spins included in the Ising model that transforms the problem to be calculated, the state is The state of the second neuron is based on the updated direction of state change of the first neuron and the weighting coefficient indicating the magnitude of the interaction between the first neuron and the second neuron. , A calculation circuit that calculates the two energy changes of the Ising model by changing 2 ⁿ (n is an integer of 0 or more) in both positive and negative directions.
Whether to allow the update of the state of the second neuron that causes the two energy changes based on the magnitude relationship between the thermal excitation energy determined based on the random number and the temperature parameter and the two energy changes, respectively. A state transition determination circuit that determines whether or not to allow the update, outputs the determination result of whether or not to allow the update, and restricts the update when the state of the second neuron exceeds the upper limit value or falls below the lower limit value.
And a plurality of Ising devices, each of which has and outputs the determination result for the second neuron different from each other.
Based on the determination result output by each of the plurality of Ising devices, a renewal neuron selection circuit that selects one of the second neurons that allow renewal and selects the state change direction, and a renewal neuron selection circuit.
A state update circuit that updates the state of the second neuron selected by the update neuron selection circuit based on the state change direction, and a state update circuit.
Optimizer with. In the state transition determination circuit, when the state of the second neuron exceeds the upper limit value or falls below the lower limit value by changing the state of the second neuron in the positive or negative direction by the 2 ⁿ , among the two energy changes. The first energy change due to a change in the state of the second neuron that exceeds the upper limit value or falls below the lower limit value is changed to a predetermined positive value, and then the first energy change is brought about. Determining whether to allow the state update of 2 neurons,
The optimization device according to claim 1. The state transition determination circuit causes the two energy changes when the state of the second neuron exceeds the upper limit value or falls below the lower limit value by changing the state of the second neuron in the positive or negative direction by the 2 ⁿ . Among the determination results of whether or not to allow the update of the state of the second neuron, whether or not to allow the update due to the change in the state of the second neuron that exceeds the upper limit value or falls below the lower limit value. The first judgment result is changed to a value indicating that the update is not allowed.
The optimization device according to claim 1. In the calculation circuit, when the state of the first neuron changes by 2 ⁿ in the positive or negative direction, the weighting coefficient having a predetermined bit width is left-shifted by the number of n bits. Use to calculate the two energy changes,
The optimization device according to any one of claims 1 to 3. It has a control unit that individually sets the value of n for each of the plurality of neurons.
The optimization device according to any one of claims 1 to 4. Of the multiple neurons whose state is represented using variables of m value (m is a positive integer of 3 or more) corresponding to multiple spins included in the Ising model that transforms the problem to be calculated, the state is The state of the second neuron is based on the updated direction of state change of the first neuron and the weighting coefficient indicating the magnitude of the interaction between the first neuron and the second neuron. , A calculation circuit that calculates the two energy changes of the Ising model by changing 2 ⁿ (n is an integer of 0 or more) in both positive and negative directions.
Whether to allow the update of the state of the second neuron that causes the two energy changes based on the magnitude relationship between the thermal excitation energy determined based on the random number and the temperature parameter and the two energy changes, respectively. A state transition determination circuit that determines whether or not to allow the update, outputs the determination result of whether or not to allow the update, and restricts the update when the state of the second neuron exceeds the upper limit value or falls below the lower limit value.
And a plurality of Ising devices, each of which has and outputs the determination result for the second neuron different from each other.
Based on the determination result output by each of the plurality of Ising devices, a renewal neuron selection circuit that selects one of the second neurons that allow renewal and selects the state change direction, and a renewal neuron selection circuit.
A state update circuit that updates the state of the second neuron selected by the update neuron selection circuit based on the state change direction, and a state update circuit.
For optimizers that have
The control device sets the weighting factor, the upper limit value, and the lower limit value.
The control device controls the magnitude of the temperature parameter.
How to control the optimizer. Of the multiple neurons whose state is represented using variables of m value (m is a positive integer of 3 or more) corresponding to multiple spins included in the Ising model that transforms the problem to be calculated, the state is The state of the second neuron is based on the updated direction of state change of the first neuron and the weighting coefficient indicating the magnitude of the interaction between the first neuron and the second neuron. , A calculation circuit that calculates the two energy changes of the Ising model by changing 2 ⁿ (n is an integer of 0 or more) in both positive and negative directions.
Whether to allow the update of the state of the second neuron that causes the two energy changes based on the magnitude relationship between the thermal excitation energy determined based on the random number and the temperature parameter and the two energy changes, respectively. A state transition determination circuit that determines whether or not to allow the update, outputs the determination result of whether or not to allow the update, and restricts the update when the state of the second neuron exceeds the upper limit value or falls below the lower limit value.
And a plurality of Ising devices, each of which has and outputs the determination result for the second neuron different from each other.
Based on the determination result output by each of the plurality of Ising devices, a renewal neuron selection circuit that selects one of the second neurons that allow renewal and selects the state change direction, and a renewal neuron selection circuit.
A state update circuit that updates the state of the second neuron selected by the update neuron selection circuit based on the state change direction, and a state update circuit.
For optimizers that have
The weighting coefficient, the upper limit value and the lower limit value are set, and the weighting coefficient, the upper limit value and the lower limit value are set.
Controlling the magnitude of the temperature parameter,
A control program for an optimizer that causes a computer to perform processing.

Images